Technology solutions powered by artificial intelligence (AI) have become mission-critical for today’s travel companies. In this roundtable discussion, FLYR executives address how AI technologies are evolving and what travel executives need to know to prioritize their investments for 2024.
This sponsored content was created in collaboration with a Skift partner.
As the year of artificial intelligence (AI) in travel continues, travel companies need to rethink their strategies around implementing AI so they can respond to customer needs and drive revenue. But before they can deploy AI systems powered by machine learning, they need to understand the technology.
Our last article in this series asked travel executives to “trust the technology,” but this article aims to provide a deeper understanding of how the technology actually works — exploring how deep learning and machine learning systems can react to situations quickly, continuously explore and exploit market opportunities, and learn from mistakes.
SkiftX invited Gene Drabkin, chief technology officer at FLYR, and Bertan Altuntas, vice president of data science at FLYR, to participate in a roundtable discussion about how AI works to help educate travel executives about potential opportunities that can be unlocked with this powerful technology.
SkiftX: What are the building blocks for travel companies to create an effective AI tech stack that can scale to meet business needs?
Bertan Altuntas: Data is the critical starting piece for making AI and machine learning successful in any organization. A good tech stack needs to ensure the data originating from different internal business processes is properly stored, managed, and distributed through the system.
Travel companies traditionally thought that data was just a byproduct and that the application system was what really mattered, so a lot of effort went into making sure applications worked well. Today, companies need to pay more attention to data quality, making data available in a timely manner, and treating data as a product.
Over the past several years, we’ve developed a better understanding of all the activities necessary at an enterprise level to make AI algorithms and systems work. For example, we know data needs to flow through multiple pipelines for algorithms to be designed properly, so we’ve collectively developed systems to observe and maintain data flows.
Data flows are made up of mathematical equations and functions. Those then become features, which are input into models. Companies need to maintain version control for these features and models and make sure the process is auditable and trackable, just as they would with organizing code in classic software development.
Experimentation is also very important. To be successful, you need an integrated development environment, just like a software engineer would have, that allows data scientists to quickly experiment with these different features, models, and algorithms and find the ones that offer the best solutions to business challenges.
Gene Drabkin: In addition to these technical systems, you need experts who can provide a gut check on what makes sense, especially within the travel and airline sectors. You need airline analysts who understand how airlines price their inventories. Another crucial element of success is bringing in AI experts from other industries who can provide the scalability and systems approach described by Bertan.
SkiftX: Can you explain how deep learning and machine learning systems work? How are they able to react to situations quickly, continuously explore and exploit market opportunities, and learn from their mistakes?
Drabkin: With machine learning, computer systems use algorithms and statistical models to draw inferences from patterns in data. Deep learning, a subset of machine learning, is essentially an artificial neural network inspired by the human brain.
Deep learning models can recognize complex patterns in data to produce accurate insights and predictions. You provide the rules, and deep learning systems look for the best possible outcome given the rewards you set up front for the model to evaluate.
Altuntas: It’s also helpful to look at the big difference between machine learning and traditional computer programming, where you need to define every detail upfront and give specific instructions to the computer on what you need to accomplish. By contrast, machine learning works by defining a meta-learning method and algorithm, and then the computer figures out how to accomplish the goal on its own.
In the context of revenue management, an analyst can define pricing rules without deep learning. For example, they might say, “Thirty days before departure, if the load factor is over 90 percent, I want the system to increase the price by 30 percent.” These rules are based on the analyst’s intuition and understanding of what happened in that market over time.
On the other hand, deep learning applications look at inputs like load factor, days before departure, capacity, and current revenue — and then calculate what the prices should be based on all of these inputs. The system is able to dynamically choose those factors and thresholds without an analyst. You can apply the model to different markets and scale up to whatever degree that you want because these models have the capacity to learn, which is truly amazing.
Another example is new modern retailing systems that utilize offer and order management. Deep learning applications can be used to segment customers based on buying history, search history, origin and destination, date and time of travel, and more, using all of these factors to determine the most relevant flights, prices, and bundles to offer the customer to boost conversion and lower acquisition costs.
Drabkin: The more relevant the data you can feed into these models that correlate to the outputs you’re looking for, the better those models will be. For example, you can add in events data — say there’s a Google Conference in San Francisco that will lead to an influx of people. The system will know that increased demand from the event will drive the price up. Other variable data points that impact price include weather and currency exchange rates.
SkiftX: An important recent evolution of AI has been improved explainability — that is, being able to “show its work” and explain how it arrived at certain decisions. Can you explain how AI is “showing its work” in the context of pricing and forecasting decisions?
Altuntas: Explaining decisions is an important way to build trust, so we are transparent about the information going into our models — our algorithms are able to show the ranking of features within a model with respect to a decision. When we get a certain pricing decision, we can see the attribution of each feature to that pricing decision.
Drabkin: For example, you can see that 60 percent of the pricing decision for an airline seat was based on competitor pricing, and 20 percent was based on historical data. It’s less of a black box.
SkiftX: How do you see AI technologies evolving over the next few years?
Drabkin: There’s a lot happening with generative AI, but reinforcement learning is better suited to creating dynamic pricing strategies. FLYR is at the forefront of reinforcement learning.
Airlines’ current approach relies on historical data, but reinforcement learning models will learn how to run by themselves the same way autonomous cars can drive by themselves — simulators teach them how to react to new things they don’t have direct experience with. This is the future of airline commercial operations.
SkiftX: What are the major hurdles that need to be overcome in the next few years to realize that vision?
Drabkin: There’s no historical data or sample data to verify against. You must simulate various scenarios to see how the model reacts, and by running multiple scenarios, the model will learn.
One challenge in airline operations is limited inventory. There’s a fixed number of seats on flights, the booking window is long, and external variables can drive the price up and down. Modeling that in a simulated environment is very difficult.
Altuntas: Defining the metric or the evaluation criteria is the hardest thing in this space because we don’t really know what’s optimal without seeing the entire market. Without understanding true optimal revenue, you need to experiment with proxy metrics, and in order to experiment and do reinforcement learning on a continual basis, you need buy-in from leadership. Not everybody wants to risk revenue in the name of learning, but if you don’t take those risks, you’ll never explore other policies that might get you to better revenue performance.
SkiftX: Do you see a role for generative AI in FLYR’s products and services?
Drabkin: On the pricing side, deep learning models are more applicable. However, one thing that benefits us from this year’s explosive buzz around generative AI is that more people are investing in AI on the supply side. For example, chip makers are creating faster and better graphic processing units (GPUs), which helps us run our deep learning models faster, cheaper, and easier.
Altuntas: With deep learning models running at faster speeds, they can evaluate complex data sets that involve hundreds of inputs, all at a very large scale that a human cannot process. That’s the value our system brings to the market — for airlines, hotels, car rental companies, and any company operating with a perishable inventory.
Drabkin: FLYR started with airlines because it’s a much more complicated space, with more variables, more players, and more historical contexts. We knew if we could solve pricing and availability problems for airlines, we could solve them for any other travel and transportation vertical.
For more information about FLYR and its commercial intelligence and business optimization platform that leverages AI and deep learning, click here.
Have a confidential tip for Skift? Get in touch